On Wed, Dec 11, 2024 at 04:50:39PM +0000, Matthew Wilcox wrote:
On Wed, Dec 11, 2024 at 11:09:56AM -0500, Johannes Weiner wrote:
This would work, but it seems somewhat complicated. The atomics in memcg charging and the vmstat updates are batched, and the per-page overhead is for the most part cheap per-cpu ops. Not an issue per se.
OK, fair enough, I hadn't realised it was a percpu-refcount. Still, we might consume several batches (batch size of 64) when we could do it all in one shot.
Perhaps you'd be more persuaded by:
(a) If we clear __GFP_ACCOUNT then alloc_pages_bulk() will work, and that's a pretty significant performance win over calling alloc_pages() in a loop.
(b) Once we get to memdescs, calling alloc_pages() with __GFP_ACCOUNT set is going to require allocating a memdesc to store the obj_cgroup in, so in the future we'll save an allocation.
Your proposed alternative will work and is way less churn. But it's not preparing us for memdescs ;-)
We can make alloc_pages_bulk() work with __GFP_ACCOUNT but your second argument is more compelling.
I am trying to think of what will we miss if we remove this per-page memcg metadata. One thing I can think of is debugging a live system or kdump where I need to track where a given page came from. I think memory profiling will still be useful in combination with going through all vmalloc regions where this page is mapped (is there an easy way to tell if a page is from a vmalloc region?). So, for now I think we will have alternative way to extract the useful information.
I think we can go with Johannes' solution for stable and discuss the future direction more separately.